A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 03:30 UTC
[Refactor] Remove Unused Env `VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON` (#20334)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Jul 1, 2025 at 22:03 UTC
9dae7d46bfce1b6b6e83e36dd6297c2c9fc58736
2 files changed
-8
vllm/envs.py
-7
@@ -104,7 +104,6 @@ if TYPE_CHECKING:
104
VLLM_SERVER_DEV_MODE: bool = False
105
VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
106
VLLM_MLA_DISABLE: bool = False
107
- VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON: bool = False
107
VLLM_RAY_PER_WORKER_GPUS: float = 1.0
108
VLLM_RAY_BUNDLE_INDICES: str = ""
109
VLLM_CUDART_SO_PATH: Optional[str] = None
@@ -769,12 +768,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
768
"VLLM_MLA_DISABLE":
769
lambda: bool(int(os.getenv("VLLM_MLA_DISABLE", "0"))),
770
772
- # If set, vLLM will use the Triton implementation of moe_align_block_size,
773
- # i.e. moe_align_block_size_triton in fused_moe.py.
774
- "VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON":
775
- lambda: bool(int(os.getenv("VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON", "0"))
776
- ),
777
-
771
# Number of GPUs per worker in Ray, if it is set to be a fraction,
772
# it allows ray to schedule multiple actors on a single GPU,
773
# so that users can colocate other actors on the same GPUs as vLLM.
vllm/model_executor/layers/fused_moe/moe_align_block_size.py
-1
@@ -94,7 +94,6 @@ def moe_align_block_size_stage4(
94
95
# Triton implementation based on:
96
# https://github.com/sgl-project/sglang/commit/ba5112ff691d791a9e38c6c71f59324a5fcb49d0
97
-# TODO(wentao): Deprecated this function in the future.
97
def moe_align_block_size_triton(
98
topk_ids: torch.Tensor,
99
num_experts: int,